Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Teaching Simulation Roger Grinde, [email protected] University of New Hampshire Files: http://pubpages.unh.edu/~rbg/TMS/TMS_Support_Files.html Teaching Simulation Do you teach simulation? In which courses? With spreadsheets? Add-Ins? Monte Carlo? Discrete Event? Do you use simulation to help teach other topics? Do other courses at your school use simulation? Session Overview Common Student Misunderstandings Simulation-Related Learning Goals Motivations Building on Other Methodologies Effects of Correlation Interpreting Results Software Issues Considerations, Recommendations Student Misunderstandings What are some misunderstandings students have about decision-making in the face of uncertainty? What are some common errors students make in simulation? Some Considerations Decide which learning goals are most important, and structure coverage so those goals are attained. Student backgrounds Time constraints Overall course objectives Inter-course relationships, role of course in curriculum Monte-Carlo and/or Discrete-Event? Related software selection question. Teaching environment, class size, TA support, etc. Learning Goals What are your learning goals when teaching simulation? Fundamental Concepts Methodology of Simulation Applications of Simulation Modeling Knowledge & Skills Critical & Analytical Thinking Simulation, Statistidcal, Spreadsheet Modeling, Decision Making Concepts Mapping: Learning Goals to Examples Examples/Application Areas X X X X X X X X X X X X X X X X X X X X X X X X X X X X Games/Tournaments X X Personal Financial Planning Multiple Project Selection Stock Price Modeling, Option Pricing Inventory (multiperiod) X X X X Queuing X Capital Project NPV Extension of other analaysis tools Is simulation needed? Variety of probability distributions Model-building issues (where a simulation model would be different than a deterministic model) Output distribution as function of input distributions Historical/empirical data Summary statistics Alternate decision criteria & risk measures Sources of error Correlation and/or relationships among input variables Optimization concepts in simulation Portfolio Allocation Learning Goal/Objective Inventory (singleperiod) Mapping: Goals to Examples X X X X X X X X X X X X X X X X X X X X X X Motivations (Why is simulation useful?) Two investment alternatives A: Invest $10,000. B: Invest $10,000 Probability of a $100,000 gain is 0.10 Probability of a $10,000 loss is 0.90 Probability of a $500 gain is 1.0 Which would you choose? Why? Risk-Informed Decision Making Appropriate and inappropriate uses of averages. Managers manage risk. Simulation gives us a tool to help us evaluate risk. Risk: The uncertainty associated with an undesirable outcome. Risk is not the same as just being uncertain about something, and is not just the possibility of a bad outcome. Risk considers the likelihood of an undesirable outcome (e.g., the probability) as well as the magnitude of that outcome. “Flaw of Averages” (Sam Savage) Article by Sam Savage (http://www.stanford.edu/~savage/faculty/savage/) Annuity Illustration (historical simulation) Simulation Model Schematic Fixed (Known) Inputs Random (Uncertain) Inputs Simulation Model Outputs & Performance Measures Decision Variables Concept of an output “distribution.” Foundations of Simulation Randomness, Uncertainty Probability Distributions Tools Dice Roller (John Walkenbach: http://www.j-walk.com/ss) Die Roller (modified) Interactive Simulation Tool Extending Other Methodologies Spreadsheet Engineering Base Case Analysis What-If Analysis, Scenario Analysis Critical Value Analysis Sensitivity Analysis Simulation Extending Other Methodologies Familiar Example/Case; Students have already developed model and done some deterministic analysis. Students provided with some probability distribution information Develop comfort with mechanics of simulation See the “value added” of simulation Provides entry point for discussion of important questions Example: Watson Truck Adapted from Lawrence & Weatherford (2001) Students have previously built basecase model, done “critical value” analysis (using Goal Seek), and have done sensitivity analysis (data tables, tornado charts) Link to files: PDF, Sensitivity, Simulation Watson Truck: Inputs A 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 B C D E F G Watson Truck Rental Parameters/Uncontrollable Inputs Purchase Price Property Tax Var Cost/Truck # Trucks Prop Tax Growth Truck Cost Growth Base Truck Rental Rate % Trucks Rented @ $1000 Rental Rate Slope Rental Rate Inflation Business Sale Multiplier Discount Rate $1,000,000 $35,000 $4,800 50 4% 7% $1,000 60% 7% 9% 3 10.0% Decision Variable Rental Rate (decision variable) $1,000 Intermediate Calculations Rental Rate Slope % Trucks Rented -0.07% 60.0% per year per year per month per $100 reduction in rental rate Sales price assumed to be 3*(year 3 revenues) per $1 increase in rental rate Watson Truck: Base Case Model 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 A B Primary Output Net Present Value (@ discount rate) C D E F G H C25: =NPV(C15,D40:F40)+C40 $209,769 Cash Flow Model 0 Cash Inflows Truck Rental Income Business Sale Total Inflows Cash Outflows Purchase Price Property Tax Truck Var Cost Total Outflows 1 2 3 $360,000 $392,400 $360,000 $392,400 $427,716 $1,283,148 $1,710,864 $1,000,000 $35,000 $240,000 $275,000 $36,400 $256,800 $293,200 $37,856 $274,776 $312,632 ($1,000,000) $85,000 $99,200 $1,398,232 F30: =E30*(1+$C13) $0 $1,000,000 F31: =C14*F30 F32: =SUM(F30:F31) F36: =E36*(1+$C8) F37: =E37*(1+$C9) F38: =SUM(F35:F37) F40: =F32-F38 Net Cash Flow Watson Truck: Sensitivity Analysis Tornado Sensitivity Chart Output Measure $0 $50,000 $100,000 $150,000 $200,000 $250,000 $300,000 $350,000 $400,000 $450,000 $500,000 Base Truck Rental Rate % Trucks Rented @ $1000 # Trucks Parameter Purchase Price Business Sale Multiplier Var Cost/Truck Discount Rate Rental Rate Inflation Rental Rate Slope Property Tax Truck Cost Growth Prop Tax Growth -10 Pct +10 Pct Watson: Simulation Forecast: Net Present Value (@ discount rate) 1,000 Trials Frequency Chart 1,000 Displayed .035 35 .026 26.25 .018 17.5 .009 8.75 Mean = $232,119 .000 0 ($525,250) ($139,570) $246,111 $631,792 Certainty is 82.00% from $0 to +Infinity Dollars $1,017,472 Learning Goals Addressed (at least partially) Linkage with other course/functional area What inputs should we simulate? Useful probability distributions. Choice of parameters. Subjective versus objective estimates. Concept of an output distribution What results are important? Sources of error in simulation Simulation mechanics Simulation in context with other tools Example: Single-Period Portfolio Simple example, but helps address a number of learning goals Do we need to simulate? Effect of correlation among input quantities Confidence vs. Prediction (certainty) intervals Quantification of risk, multiple decision criteria Optimization concepts within simulation context Precision of estimates from simulation Link to file Spreadsheet B 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 C D E Portfolio Allocation Model Investments Money Market fund Income fund Growth and Income fund Aggressive Growth fund Total amount available Annual return 3.0% 5.0% 7.0% 11.0% $100,000 Decision variables Money Market fund Income fund Growth and Income fund Aggressive Growth fund Total expected return Amount invested $25,000 $25,000 $25,000 $25,000 $6,500 Lower bound $0 $10,000 $0 $10,000 Upper bound $50,000 $25,000 $80,000 $100,000 Do we need simulation? Assuming we know the distributions for the returns, do we need simulation to compute the expected return of the portfolio? variance of the portfolio? tail probabilities? What if the asset returns are correlated? Large Stocks Large Growth Stocks Large Value Stocks Small Stocks Small Growth Stocks Small Value Stocks Foreign Stocks Bonds What is the effect of correlation on the distribution of portfolio returns? Large Stocks Large Growth Stocks Large Value Stocks 1 0.958411 0.901159 0.720036 0.755265 0.507037 0.391551 0.366054 1 0.74063 0.606356 0.720758 0.322101 0.294704 0.267404 1 0.776748 1 0.678815 0.921266 1 0.715396 0.875155 0.618755 1 0.454882 0.275857 0.325624 0.140232 1 0.465424 0.28663 0.164181 0.369522 0.112362 Small Stocks Small Growth Stocks Small Value Stocks Foreign Stocks Based on Standard & Poor Micropal, via Franklin/Templeton Investor Topics Update, Winter 2001 (Asset Returns from 1980-2000) Bonds 1 Results (n=1000) No Correlation Mean = $6842 Standard Deviation = $5449 5% VaR = ($2165) Positive Correlation Mean = $6409 Standard Deviation = $7386 5% VaR = ($5655) Decision Criteria, Risk Measures What criteria are important for making decision as to where to invest? Average? Standard Deviation? Minimum? Maximum? Quartiles? VaR? Probability of Loss? Measures of risk. Simulation gives us the entire output distribution. Entry point for optimization within simulation context Alternate scenarios, efficient frontier, OptQuest, RiskOptimizer, etc. Confidence Intervals Students can (usually) calculate a confidence interval for the mean. Do they know what it means? Reconciling confidence and prediction intervals. Sample Results (Portfolio Problem) Statistics: Trials Mean Median Mode Standard Deviation 90% Confidence Interval Standard Error $233.56 Z 1.645 Lower Limit $6,025 Upper Limit $6,794 90% CI on Mean Dollar Return: ($6025, $6794) What does that confidence interval mean? Value 1000 $6,409 $6,531 --$7,386 Common (student) error What does the CI about an individual outcome? For example, from this year’s return? Sample Results (cont) Percentile 0% 5% 10% 15% 20% 25% 30% 35% 40% 45% 50% 55% 60% 65% 70% 75% 80% 85% 90% 95% 100% dollars ($16,088) ($5,655) ($3,052) ($1,008) $271 $1,277 $2,498 $3,484 $4,307 $5,365 $6,531 $7,695 $8,349 $9,310 $10,386 $11,419 $12,431 $13,896 $15,689 $18,659 $30,330 Cumulative Percentiles of the Portfolio Return Distribution What do these results mean? What is the 90% “prediction” (or “certainty”) interval (centered around the median)? Putting Them Together 90% Confidence Interval for the Mean 90% Prediction Interval (centered around median) ($6025, $6794) (-$5655, $18,659) Note: Crystal Ball uses the term “certainty”) Students: Understand the difference? Understand when one is more appropriate than the other? Precision of Simulation Results Since we know the true value of the mean (for the portfolio problem), this can be a good example to look at precision and sample size issues. Confidence interval for proportion or for a given percentile sometimes makes more sense. Crystal Ball: Precision Control Nice way to illustrate effect of sample size. Precision Control stops simulation based on user-specified precision on the mean, standard deviation, and/or a percentile. Actually, CB stops whenever the first of a number of conditions occurs (e.g., maximum number of trials, precision specifications). Example (Portfolio Allocation) Example (Option Pricing) Precision: Portfolio Example Precision Control Experiment: Summary Number of Trials Required for Specified Precision Precision $800 $400 $200 $100 Mean 300 1250 5350 21,700 5th Percentile 1900 4950 20,300 95,200 Note: "Precision" is the half-width of the 95% confidence interval True Mean = $6,500 Precision: Option Pricing Example Precision Control Experiment: Mean Call and Put Values Number of Trials Required for Specified Precision of Mean Call and Put Values Trials 50 200 700 2600 64400 Precision: Call Precision: Put $0.259 $0.390 $0.166 $0.181 $0.099 $0.090 $0.050 $0.047 $0.010 $0.009 Note: "Precision" is the half-width of the 95% confidence interval Black-Scholes Call Price = $0.733 Black-Scholes Put Price = $0.955 Crystal Ball Functions and Simple VBA Control Crystal Ball provides built-in functions Control through VBA Distribution Functions (e.g., CB.Normal) Functions for Accessing Simulation Results (e.g., CB.GetForeStatFN) For some students, can be a hook into greater interest in simulation and/or VBA/DSS. Allows one to prepare a simulation-based model for someone who doesn’t know Crystal Ball. Example VBA-Enabled Example B 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 C D E F G H I J Portfolio Allocation Model Investments Money Market fund Income fund Growth and Income fund Aggressive Growth fund Total amount available Annual return 3.5% 4.1% 6.5% 29.3% $100,000 Decision variables Money Market fund Income fund Growth and Income fund Aggressive Growth fund Total return Amount invested $25,000 $25,000 $25,000 $25,000 $10,866 Param1 2.0% 5.0% 7.0% 11.0% Distributions/Parameters Param2 Distribution 4.0% Uniform 5.0% Normal 12.0% Normal 18.0% Normal C5: =CB.Uniform(D5,E5) C6: =CB.Normal(D6,E6) (copied down) $C$17: =SUMPRODUCT(C5:C8,C13:C16) (Forecast Cell) C22: =CB.GetForeStatFN(C$17,1) Simulation Results Summary Statistics Number Trials Mean Standard Deviation Minimum Maximum Standard Error of Mean 400 $6,695 $5,644 ($9,718) $20,612 $282 Value at Risk (enter %) Critical Value (enter $) 5% $4,000 D30: =CB.GetCertaintyFN(C$17,C30)/100 Quartiles 0% 25% 50% 75% 100% ($2,985) 0.305 ($9,718) $2,840 $6,931 $10,730 $20,612 5.0% chance return will be <= ($2,985) Probability (Return <= $4,000) = 0.305 F22: =CB.GetForePercentFN(C$17,E22*100) Enter Number of Trials 400 Run Simulation CB. Functions and VBA CB. Distribution Functions CB. Functions for reporting results e.g., CB.Normal, CB.Uniform, CB.Triangular) CB.GetForeStatFN, CB.GetCertaintyFN, CB.GetForePercentFN VBA: simple to automate specific processes Sub RunSimulation() CB.ResetND CB.Simulation Range("n_trials").Value End Sub Sub CreateReport() CB.CreateRpt ' CB.CreateRptND cbrptOK End Sub Learning Goals Revisited Decide which learning goals are the most important, and structure coverage so those goals are attained. Student backgrounds Time constraints Overall course objectives Mapping of learning goals to examples, cases, and projects that you will use. Simulation, Statistidcal, Spreadsheet Modeling, Decision Making Concepts Mapping: Learning Goals to Examples Examples/Application Areas X X X X X X X X X X X X X X X X X X X X X X X X X X X X Games/Tournaments X X Personal Financial Planning Multiple Project Selection Stock Price Modeling, Option Pricing Inventory (multiperiod) X X X X Queuing X Capital Project NPV Extension of other analaysis tools Is simulation needed? Variety of probability distributions Model-building issues (where a simulation model would be different than a deterministic model) Output distribution as function of input distributions Historical/empirical data Summary statistics Alternate decision criteria & risk measures Sources of error Correlation and/or relationships among input variables Optimization concepts in simulation Portfolio Allocation Learning Goal/Objective Inventory (singleperiod) Mapping: Possible Learning Goals to Examples X X X X X X X X X X X X X X X X X X X X X X Common Student Errors Thinking of simulation as the method of first choice. Simulating too many quantities. Too much focus on distribution/parameter selection or on the numerical results, not enough on insights/decision. Misinterpretation of results, especially confidence intervals Modeling: Using same return, lead time, etc. for every time period/order, etc. (difference between deterministic and simulation models) Choosing the assumptions, distributions, parameters, etc. that give the “best” numerical results. Software Issues: Monte-Carlo Alternatives “Full-Service” Add-In? (e.g., @Risk, Crystal Ball, XLSim by Sam Savage, RiskSim) “Helper” Workbook? (e.g., Interactive Simulation Tool with Random Number Function support) “Native” Excel? All have advantages, disadvantages Back to learning objectives, role of course, student audience, etc. Software Issues: DiscreteEvent Alternatives Stand-alone package (e.g., Arena, Process Model, Extend) Excel Add-In (e.g., SimQuick by David Hartvigsen) Native Excel modeling augmented by Monte Carlo tool (e.g., QueueSimon by Armann Ingolfsson) DE Simulation can be a great way to help teach concepts in other areas (e.g., queuing, inventory) Don’t necessarily need to teach DE Simulation to be able to use it to teach other things. Other Considerations Program-level, inter-course objectives Role of course in curriculum Level/background of students Monte-Carlo and/or Discrete-Event? Related software selection question. Teaching environment, class size, TA support, etc. How much of course can/should be devoted to simulation? Recommendations Learning Goals: Figure out what you really want students to learn and be able to do, after your class is over; in other classes, internships, future jobs? How can simulation coverage help accomplish these goals? Cases: Engage students in the business problem, let them discover relevance of simulation. Student-Developed Projects: Students gain better awareness of all the “little” decisions involved in modeling and simulation. Additional Slides Concept Coverage Through Examples Philosophy: Expose students to a number of application areas, but at the same time covering fundamental decision-making, modeling, and analysis concepts and methodologies. Counter to the way many of us were taught. Key: We need to clearly understand which concepts we’re trying to convey with each example. Examples that Work Well Fundamentals: Dice Roller, Interactive Simulation Tool Personal Decisions: Car Repair/Purchase Decision, Portfolio (single period, based on CB Model), College Funding (based on Winston & Albright) Capital Project Evaluation: Truck Rental Company (based on Lawrence & Weatherford), Project Selection/Diversification (CB Model), Product Development & Launch (CB Model) Finance: Stock Price Models, Option Pricing, Random Walks, Mean Reverting Processes Examples (continued) Inventory: DG Winter Coats (NewsVendor), Antarctica (multi-period, based on Lapin & Whisler) Queuing: QueueSimon (Armonn Ingolfsson) Games/Tournaments, Sports: NCAA Tourney (based on Winston & Albright), Home Run Derby Baseball Simulation (VBA-enabled), Baseball Inning Simulation Simulation in Teaching Other Topics: Revenue Management Illustration, QueueSimon (Armonn Ingolfsson) Crystal Ball Features: CB Macros, CB Functions Examples Posing Difficulties for Spreadsheets Multi-server queues and queue networks Most production systems Business process redesign However, some add-ins do exist for simple discrete-event models (e.g., SimQuick by David Hartvigsen) Sources of Error in Simulation What are some of the sources of error in a spreadsheet simulation model/analysis? Learning Objectives (Revisited) General Probability Distributions Statistics Relationships Among Variables Decision Making Possible Learning Goals General Use simulation as an extension of other analysis tools Apply simulation to a variety of business problems Identify when simulation is and is not needed to analyze a situation Probablilty Distributions Understand and use probability distributions to model phenomena Describe the output distribution, understanding this to be a function of the input distributions Use historical/empirical data and subjective assessments appropriately in choosing distributions and parameters Possible Learning Goals (cont) Statistics Correctly interpret summary statistics, including percentiles/histograms Correctly interpret confidence and prediction (certainty) intervals Identify sources of error in simulation, apply to specific situations Relationships Among Variables Include appropriate correlation and/or other relationships when model building Describe the effect of correlation and/or other relationship on simulation results Possible Learning Goals (cont) Decision Making Identify and correctly use different risk measures Use appropriate criteria in making recommendations Use optimization concepts in a simulation application Student Project Example (MBA) PPT File Excel File